Purpose: To develop a deep learning-based diagnostic system using MRI to automatically segment vertebral metastases and identify their primary origin.

Methods: MRI data from patients with spinal metastases were retrospectively reviewed at two institutions. Data from Institution I (October 2008 to March 2025) were used for model development, while data from Institution II (December 2012 to March 2025) served as an independent external test set for model evaluation. The segmentation model was developed using the Swin-UNETR architecture, integrated with a hybrid active learning strategy (HALS). A subset of cases was randomly selected from Institution I for initial model training. Then, using HALS, the Swin-UNETR model from the previous round selected the most informative samples for annotation, adding 30 new cases in each of five rounds. Segmentation performance was evaluated using the Dice coefficient. The segmented volumes of interest, along with basic demographics, were fed into the Wavelet and Attention-based Multi-modal Network (WAMNet) to identify primary sites. WAMNet’s performance was assessed using standard accuracy, weighted precision, recall, F1 score, specificity, and micro-AUC at the Top-1, Top-2, and Top-3 levels. Additionally, WAMNet’s performance was compared to that of the ResNet50 model.

Results: A total of 1,448 patients (mean age 61 ± 11 years; 675 females) with vertebral metastases from six primary sites (breast, prostate, liver, lung, kidney, and gastrointestinal) were included. To develop the classification model, all cases from Institution I were randomly split into a training set (n = 751), a validation set (n = 213), and an independent internal test set (n = 112) in a 7:2:1 ratio. All cases (n=372) from Institution II served as an independent external test set for both models' evaluation. Initially, 224 cases were manually annotated to develop the segmentation model. Among the three MRI sequences, the pretrained model performed best on T1WI, achieving a DSC of 0.55. Over five rounds of HALS-assisted sampling, the Swin-UNETR segmentation model was ultimately trained on 374 cases, improving the Dice coefficient from 0.55 to 0.75. For classification, WAMNet outperformed ResNet50, with higher accuracy and AUC at the Top-3 level in the internal (88% vs 77%; 80% vs 70%, p = .004) and external (87% vs 81%; 82% vs 79%, p = .08) test sets.

Conclusions: The automated diagnostic system demonstrated promising performance in identifying their primary origin. It predicts the most likely primary sites, effectively narrowing the differential diagnosis. This could serve as a valuable diagnostic aid by streamlining clinical workflows and potentially reducing the need for invasive diagnostic procedures.